Dual Contrastive Learning Network for Graph Clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Contrastive learning is popular for graph representation, maximizing mutual information between augmented graph views.
• Existing methods suffer from representation collapse, where features become too similar, reducing discriminative power.

Purpose of the Study:

• To propose a novel self-supervised learning method, the Dual Contrastive Learning Network (DCLN), to address representation collapse in graph contrastive learning.
• To enhance the discriminative capacity of graph representations by reducing redundant information.

Main Methods:

• Introduced the Dual Curriculum Contrastive Module (DCCM) to approximate node and feature similarity matrices.
• Utilized a curriculum learning strategy to manage sample imbalance during contrastive learning.

Main Results:

• The DCLN effectively reduces redundant information in learned latent variables.
• The proposed method preserves informative information from high-order neighbors while eliminating irrelevant features.
• Experiments on six benchmark datasets show superior performance compared to state-of-the-art methods.

Conclusions:

• The DCLN significantly improves the discriminative capacity of graph representations.
• The dual contrastive approach and curriculum learning strategy are effective for graph representation learning.